The result of our previous work on unsupervised coreference resolution, leads us to the conclusion that features are more important than the resolution method. In other words, with the same set of (informative) input features, different systems perform more or less the same.
In this regard, we develop a new supervised method for extracting precise coreference rules. Each resulting rule is an informative combination of attribute-values which precisely identifies a subset of coreference relations.